Examples

Example	Survey	Model answers	Shows
Toy	the paper's toy cells	models M1 to M3	equal accuracy, different fidelity
CES	Cooperative Election Study 2024, 96 cells	qwen3-vl:2b on Ollama; configurations for OpenAI, Anthropic, Google, vLLM, Hugging Face and transformers	the whole pipeline
Twin-2K-500	Twin-2K-500 wave 4, 24 cells	GPT-4.1-mini digital twins and a human retest	a proprietary model scored without an API key
GlobalOpinionQA	four World Values Survey items in 90 countries, fetched at run time	qwen3-vl:2b on Ollama, one country per interview	similarity against fidelity across countries
Machine Bias	World Values Survey cells	archived and fine-tuned models	the paper's tables recomputed

In the CES example, a 2-billion-parameter model is close to every cell on average (accuracy 0.71 to 0.90) but does not order the cells as the survey does, so structure binds and PFS stays at 0.30 or below. In Twin-2K-500, people answering the same items again reach PFS 0.95; digital twins reach 0.66 with a full persona and 0.58 with demographics only, held back by adaptability.